[
https://issues.apache.org/jira/browse/HIVE-13920?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15311812#comment-15311812
]
Gopal V commented on HIVE-13920:
--------------------------------
Preallocating might actually allocate large number of arrays without actual
need - a 10,000 row split shouldn't need 128 vector batches pre-created and
adding to the garbage.
The allocation is currently inside the actual reader inner loop, which is a
problem because for the first 128 invocations (& more), the allocation is
actually performed.
The 128x more allocations than the regular Tez codepath is the problem, moving
it around to a different part of code doesn't fix the actual perf issue.
> LLAP: pre-populate CVB cache
> ----------------------------
>
> Key: HIVE-13920
> URL: https://issues.apache.org/jira/browse/HIVE-13920
> Project: Hive
> Issue Type: Bug
> Reporter: Sergey Shelukhin
>
> To avoid allocations on the main path, we should pre-populate the CVB cache.
> The main difficulty is propagating column indexes earlier to allocate correct
> vectors.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
